The Prediction Register — Which Way Value Moves

Canonical: https://thonly.org/research/program/register · Licence: CC0 1.0

What this is

Seventy pre-registered predictions arising from the research program stated in [which-way-value-moves](which-way-value-moves.md). One further prediction is withheld from publication (operational channel economics); its existence is recorded here so the count is honest, bringing the true total to seventy-one.

This is a living document and the count rises. It opened at sixty-seven on 2026-08-27; see Revisions at the foot for what has been added since, and read the count in the Summary table rather than any sentence that repeats it elsewhere.

Status vocabulary. Unrun — registered, no observation attempted. Running — instrument live, data accumulating, not yet read. Resolved — read against its falsifier. Contradicted — the data went against it. Retired — superseded by a ruling that made it moot; kept, never deleted.

Provenance. Published — the prediction already appears in a corpus paper carrying that paper's OpenTimestamps and RFC 3161 proofs. First public here — the prediction existed only in internal records until this document; its first external timestamp is this file's.

Reading rule. Predictions marked running must not be read early, and participants must not be prompted toward the predicted behavior. A prompt voids the test.

---

Summary

	Count
Total registered	124
Published here	123
Withheld (operational)	1
Already carried by a published paper	58
Memory-only at first publication (2026-08-27)	22 (21 published here + 1 withheld)
Receiving first public timestamp here	65
Run	7 (two desk censuses · five formal checks: P-FA1, P-FA2a, P-FA2b, P-FA4, P-FA5)
Field tests run	0 — six pilot-scale predictions are Running with a baseline read; none has resolved
Running (instrument live; unread, restricted or baseline only)	9
Contradicted	1
Retired	3

---

Core-level

If one of these fails, the program ends rather than adjusts.

ID	Prediction	Falsifier	Status	Provenance
P-GO1	A system forbidding backward return produces more onward giving per unit received than an otherwise-identical permissive system	Equal or greater onward giving under the permissive condition → the constraint is aesthetic, not mechanical	Unrun	First public here
P-K1	As external subsidy for a kind act approaches zero, the rate of the act does not approach zero	Kindness rate tracks subsidy proportionally → the behavior was priced throughout	Unrun	First public here
P-L2	Witnessed presence generates more downstream giving than unwitnessed presence of equal duration	Equal downstream gratitude when unwitnessed → the chronicle is decorative and the institution's core claim is wrong	Unrun	Published (co-presence-gated-redemption §10)
P-R2	Product-dependence begins to fall after the inflection	Dependence never falls → the raft is a boat	Unrun	First public here
P-B3	The proportion of creators electing to dedicate their share does not decline over time	Dedication collapses once money is visible → the offering frame is eroding	Unrun	Published (dedicatory-generation)

---

Chapter I — Capacity

Does receiving create giving? If not, forward-only motion is depletion, not circulation.

ID	Prediction	Status	Provenance
P-IA1	The receiver is legible as benefactor, not need — the sitter supplies a merit-field rather than displaying a lack	Unrun	Published (inverted-alms-round)
P-IA2	Second inverted-alms prediction as registered in the paper	Unrun	Published (inverted-alms-round)
P-IA3	Third inverted-alms prediction as registered in the paper	Unrun	Published (inverted-alms-round)
P-G1	Once the app manufactures obligation, a return visit no longer distinguishes gratitude practice from guilt — publish either way	Unrun	Published (sacrifice-witness-without-discharge)
P-G2	Second B-Grace prediction as registered in the paper	Unrun	Published (sacrifice-witness-without-discharge)
P-G3	The meal-anchor outperforms gamification; gamifying at launch makes this untestable	Unrun	Published (sacrifice-witness-without-discharge)
P-W1	Users with higher thing-capture frequency show higher rates of the primed downstream behavior — registered pre-build	Unrun	Published (appreciation-as-world-building)
P-E1	First graduation-by-transmission prediction as registered	Unrun	Published (the-game-that-graduates-you)
P-E2	Second graduation-by-transmission prediction as registered	Unrun	Published (the-game-that-graduates-you)
P-E3	Third graduation-by-transmission prediction as registered	Unrun	Published (the-game-that-graduates-you)
P-M2	Matched-stakes prediction as registered	Unrun	Published (the-wager-that-isnt)
P-M3	Patron-pot prediction as registered	Unrun	Published (the-wager-that-isnt)
—	Kids-as-triggers — children are the primary triggers of household self-thanking	Contradicted (parents self-thanked more; n=1, confounded, not hardened)	Published (kids-as-triggers-self-thanking)
P-BG1	A shorter re-address window on a relayed gift increases issuances and decreases COMPLETED placements. Null: completions hold or rise as the window shortens — the tighter clock is free. ⚠️ Unanswerable unless issued · passed · lapsed · re-issued are logged from the first instrument	Unrun (unscorable as specified at n=1 — one ruled window, no variance; see P-BG1a)	First public here
P-BG2	Sustained pass-on is higher for an ANONYMOUS relayed gift than a VISIBLE one, over a window long enough for the effect to appear. Null: rates match or visible exceeds anonymous — the anonymity is doctrinal only, and the visible relay should be deprecated for minors. Rests on the developmental finding that adult-directed giving underperforms child-led	Unrun	First public here
P-BG3	Within a bounded family, VISIBLE gifts produce more give-BACK to the giver within 30 days; ANONYMOUS gifts produce more give-FORWARD to third parties. Null: rates match, or anonymous produces more give-back — the obligation mechanism is not operating. ⛔ Analyse PEER and ADULT-TO-CHILD strata separately — the effect is expected between peers and weak or absent in adult-to-child care, and pooling would produce a false negative. ⚠️ Tests the mechanism, not the atmosphere claim it was offered to support	Unrun	First public here
P-BG1a	Corrects P-BG1, which cannot be scored as specified — P-BG1 compares window lengths, and one re-address window (one month) is ruled for every relayed gift. Under that one window: among relayed gifts that reach a recipient, at least 80% are passed within 7 days of issuance, and the pass-time distribution shows no rise in the window's final third (days 21–30 carry no more placements than days 11–20). Equivalently, by truncation counterfactual, a 7-day window would have censored fewer than 20% of the placements observed. Falsifier: placements cluster in the final third, or a 7-day truncation censors 20% or more → the clock is doing work, and P-BG1's "a shorter window loses placements" half is live at pilot scale. Scored once 20 or more completed placements exist. Instrument: issuance-to-pass time, written per instrument by the relayed gift's append-only event log from the first gift. ⚠️ A truncation counterfactual bounds; it does not measure — it assumes no behavioural response to a shorter window, so a pass licenses the window is not obviously load-bearing, never shortening it is free. P-BG1's issuance half stays unscorable	Unrun	First public here
P-BG6	Who paid does not change pass-on — among relayed gifts that reach a recipient, the share passed on within the one-month window is the same whether Miss Aquarius paid from the Pool or an anonymous family member did: the two rates fall within 10 percentage points. Falsifier: they differ by more than 10 points, in either direction → who paid changes what the recipient does with the gift. Scored once each arm has 20 or more gifts reaching a recipient (P-BG1a's floor). Instrument: pass-on per gift from the relayed gift's event log, split by payer, which needs the payer recorded on each gift (roadmap A73). Direction and the 10-point band are the founder's	Unrun (payer not yet recorded per gift)	First public here
P-BG4	A gift code draws more gratitude than a tip — within the pilot family, among gift codes and tips a member RECEIVES worth $2 or less at the day's rate, the share whose recipient sends the giver a thank (words; a gift's thank-back door counts, an anonymous giver's too) within 7 days of receipt is at least 10 percentage points higher for gift codes than for tips. ⛔ A thank that came through a shop's thank door (marked with the shop on the thank) is excluded: it goes to a shop's staff, not to the giver, so it is not gratitude to the giver. Falsifier: the tip rate matches, exceeds, or trails by under 10 points → tangibility does not draw gratitude the way the founder predicts. Receipts from registration to 2027-01-06 (one season, before the January 7 reset); scored once 20 gift codes and 20 tips in that band exist, else unscorable — the window is not extended. Gratitude is a thank issued, never value returned (value returned is P-BG3's obligation); the instrument split keeps the two readable. ⚠️ The stated mechanism is tangibility; the founder's own most-remembered gratitude came from letters, the least tangible givable, so P-BG5 is the arm that tells tangibility from attribution. Direction is the founder's; parameters proposed and adopted by him	Unrun	First public here
P-BG5	An attributed, never-pooled tip draws gratitude equal to a gift code — the discriminating arm of P-BG4: a tip labelled with its giver and never pooled into a balance draws a thank back within 7 days at a rate within 10 percentage points of a gift code's, by P-BG4's instrument and exclusion. Falsifier: the gift code still exceeds it by more than 10 points → tangibility does independent work. If P-BG5 holds, the mechanism is attribution and P-BG4's stated cause is wrong even if P-BG4 itself passes. Unrun until a never-pooled tip exists (today every tip lands in a balance); then P-BG4's floors apply. Parameters proposed and adopted by the founder	Unrun (no never-pooled tip exists)	First public here
P-BG7	Shops beyond the family opt in — of the first five shops outside the founding family to open, at least one owner consents to Miss Aquarius crediting the owner's Fund (entry to the rotation she funds in kind) within 90 days of opening. Falsifier: none of the five consents in its first 90 days → capacity in kind reaches only shops the family already knows, and universal giving does not travel through strangers' shops. Scored when the fifth such shop is 90 days old. Parameters proposed and adopted by the founder	Unrun	First public here
P-PL1	Durability — the founding family's 28-day capture rate returns to the June 2026 baseline of 4.5 moments/week at some monthly read before 2027-01-01. Falsifier: no read reaches it → the June inflection was novelty, not a level. Window and baseline are the ones ratified for P-C0a. Read 1 of 3 (2026-09-05): 2.25/wk	Running (baseline read)	First public here
P-PL4	Pass-forward — more than half of re-tip value received (each document's own recorded USD cents) is re-given by its recipient within 30 days: Σ min(received, given) / Σ received over a 90-day receipt window that closes 30 days before the read. Falsifier: half or less → received value is income, not a licence to give; the re-patronage conduit is anecdote. The give-forward atom made numeric. Read 1 of 3: 43%	Running (baseline read)	First public here
P-PL5	Inelasticity — the pilot-scale arm of P-K1. Giving persists where the AI reward is zero: the share of givers with zero AI reward in a month does not fall as subsidy is withdrawn. ⚠️ Not scorable until the taper (scheduled Q2 2027, not to be pulled forward) and the F3QR confound registered 2026-09-04 against P-K1 are resolved; reported each read (2026-09: 86% of givers received no AI reward that month)	Unrun (instrument gated)	First public here
P-PL6	The top three recipients of re-tip value in the founding family hold under half of value received in any 28-day window with five or more recipients. ⛔ Retired in the revision that registered it: it fails by arithmetic — with five recipients an even spread already puts the top three at 60%, so it could not pass until seven had received. Superseded by P-PL6a; kept, not deleted	Retired	First public here
P-PL6a	Corrects P-PL6. Accumulates to no one — in the founding family, the top three recipients of re-tip value in a 28-day window hold less than 1.5× their share under a perfectly even spread (share ÷ min(1, 3/n)), once eight or more people have received. Falsifier: 1.5× or more → value pools around a few; the jubilee and the Re-Tip Jar are not doing their structural work. The same property freed of n. Read 1 of 3: 1.85× (8 recipients)	Running (baseline read)	First public here
P-PL7	Anonymity — anonymous givers reach more distinct recipients than visible givers (the demand-side thesis: the capacity to give to loved ones, anonymously; the pilot-scale sibling of P-BG2). ⚠️ No variance: 827 of 830 re-tips carry the anonymous flag, so the field records a default, not a choice. Unscorable until the product offers a real choice	Unrun (no variance)	First public here
P-PL8	Kids as triggers, de-confounded — enters as a new prediction citing the Contradicted finding above, under the revision rule: within the founding family, the younger generation captures more self-thanks per active member than the older once device access is controlled for. The first result was contradicted at n=1 and ruled not to be hardened; this names the confound and the instrument (the device join from the error stream, plus the interview leg)	Unrun (instrument not built)	First public here
P-PL9	Diaspora corridor, dignity direction — low-income→high-income flows (KHR sender to USD recipient, currency as an income proxy) are at least 10% of money-carrying re-tips in the trailing 90 days, once twenty or more flows exist. Falsifier: under 10% → patronage is a corridor, not a mesh; the low-income members are recipients only. June 2026 baseline 9.6%. Read 1 of 3: 6.0% of 712	Running (baseline read)	First public here
P-PL12	Cross-income gravitation, among strangers — a USD-currency viewer re-thanks a KHR-currency creator's B-Short at at least 2× the per-view rate at which they re-thank a USD-currency creator's B-Short, counting only viewer–creator pairs in different families, over a trailing 90-day window, once each arm has 500 or more qualifying views (watch-depth rows, schemaVersion 2 only) and twenty or more re-thanks exist in total. Falsifier: under 2× → the high→low flow the pilot shows is family obligation, not a preference among strangers. Per view, not share of flows, so that what the feed happens to show cannot pass it. Currency is an income proxy. ⛔ Void if any surface orders, filters or labels B-Shorts by country or income — a prompt voids the test. Direction, measure and threshold are the founder's; the sample floors are substrate-set	Unrun (awaits cross-family B-Short visibility and its per-view metric)	First public here
P-ZP1	The game travels with the foundation — a new prediction citing P-K1 and P-PL5, neither edited. Among givers active in the three months before the taper, those in the top third by Zero-Point play keep giving through the taper at at least 1.25× the rate of those in the bottom third. Zero-Point play = distinct acts of giving forward per active month: re-tips passed on (live now), thanks given after a B-Tag (live in B-Shops), and whole-balance releases in B-Promise (from its launch). Keep giving = gave at least once in each of the three months after the taper begins. Falsifier: a ratio under 1.25×, or the bottom third keeping on at the same rate or higher → Zero-Point play does not travel with the kindness the proxy measures, and the game is not evidence of the foundation. Scorable only at the P-K1/P-PL5 taper (scheduled Q2 2027, not to be pulled forward), with at least 10 givers in each third; below that it is reported, not scored. The F3QR confound registered against P-K1 on 2026-09-04 binds this too. ⚠️ What it cannot show: the generous may simply choose to play more — a pass means the game and money-inelastic giving travel together, never that the game causes it. Direction, measure and threshold are the founder's (registered as drafted)	Unrun (gated on the taper)	First public here
P-PL13	Shop gifts add to thanking — once a pilot member's first shop gift is sent or received through Treasury, the thanks they issue per month rise by more than 20% of their own pre-exposure rate beyond the change in unexposed members over the same months (a difference-in-differences on thanks issued). Exposure = the member's first shop gift order as payer or holder (Treasury keeps these records from 2026-10-05). Thanks issued = thanks sent from Treasury's own surfaces; a thank that came through a shop's thank door (marked with the shop on the thank) is excluded from the outcome and reported beside it, so a new channel cannot confirm this by itself. Falsifier: the difference is within ±20% of the pre-exposure rate, or negative → shop gifts do not add to thanking, or substitute for it. Per read; before three reads, or while fewer than five members are exposed, the value is shown and not scored (the pilot's standing floors). ⚠️ One family, and gifts arrive alongside other changes: a pass licenses gift exposure travels with more thanking, never gifts cause it. The direction is the founder's, chosen against a no change answer bounded at ±20%	Unrun (metric built 2026-10-07; no member exposed yet)	First public here
P-SD1	The thank door is used — at Home Coffee, the first HeartBank® Shop, over the 90 days from its first completed order after 2026-10-07, thanks given in Treasury through the shop's thank door (each carries the shop's mark) number at least 5% of the shop's completed orders. The door is a video thank to the owner that gives Kiitos and carries no money; it asks for a Treasury account. Falsifier: under 5% → the video requirement and the account wall keep Kiitos out of customers' reach, and "Kiitos always, cash optional" holds in the rules, not in use. Instrument: two monthly counts — Treasury's door thanks and the shop's completed orders — never joined to a person. Scored once the shop has 100 or more completed orders in the window. ⛔ The door is never promoted: a prompt voids the test.	Unrun	First public here
P-PL14	The 7 January reset moves held capacity into giving — the use-it-or-lose-it claim published in giving-is-a-gift-too, tested rather than ruled. In the final 7 days before the pilot family's first reset (2026-12-31 to 2027-01-06, Phnom Penh time), each member's share of money-thanks drawn from the Re-Thank Foundation℠ rises over that member's own baseline share (2026-10-14 to 2026-12-23), and the median member's change is at least +10 percentage points. Share = Foundation-drawn value ÷ (Foundation-drawn + Personal Account–drawn value), where a re-thank draws on the Foundation first and a thank's tip only on the Account; comparing shares, not totals, cancels proportional holiday spending. The founder's own sends are excluded (payer and observer). Falsifier, two-sided: the median change is within ±10 points → no deadline effect, the published claim not supported; −10 points or lower → a slump, members held back before the reset. Either counts against the claim. Scored once, after 2027-01-07, with at least three members active in both windows. ⚠️ One household, one season; the threshold was set by the substrate before any data and without a variance estimate; the treatment includes the app's own wording (the reset date shown beside the Foundation balance, a Khmer phrasing that reads as a deadline), deliberately left unchanged until after the reset. ⛔ Never prompted: a prompt voids the test. Occasioned by the H3 position what-an-argument-cannot-reach (In Reply No. 1)	Unrun (instrument recording since 2026-10-14 — every re-thank stores its Foundation and Account parts; the read itself is not built)	First public here

---

Chapter II — Substrate

Does an hour behave like a dollar? If not, the invariance claim is false.

ID	Prediction	Falsifier / note	Status	Provenance
P-C0a	Completed pledges produce captured moments at a rate no higher than the family's baseline capture rate	If so, the handoff is decorative and Chronicle v1 should not be built around it. Read on the 28-day window against 4.5/wk	Running, unread	First public here
P-C0b	Shorts linked to a completed time-sacrifice do not out-perform ordinary moments on watch-through and re-tip rate	⚠️ Instrument defect, disclosed: completion rate is biased against the hypothesis — longer recordings score lower by construction, so P-C0b would read confirmed mechanically. Restricted to views/wk and re-tip rate until repaired. ✅ Repaired 2026-09-05: views rows written from that date carry real watch depth (schemaVersion 2); earlier rows are never pooled with them, so the restriction lifts only for post-repair rows	Running, restricted	First public here
P-C0c	Redeemers who capture cannot articulate the re-tip motive when asked afterward	Tests whether the stated mechanism arises unprompted, which neither original falsifier covered	Running, unread	First public here
P-C1	Chronicle handoff prediction as registered	—	Unrun	Published (co-presence-gated-redemption)
P-C2	De-facto-rate tripwire, publish either way	—	Unrun	Published (co-presence-gated-redemption)
P-C3	Standing de-facto-rate tripwire, publish either way	—	Unrun	Published (co-presence-gated-redemption)
P-L1	Hours co-spent predict relationship warmth better than money sent, messages exchanged, or stated affection	—	Unrun	Published (co-presence-gated-redemption)
P-L3	Two arms — silent after pledge vs one offer to reach out — measuring redemption rate and recipient-reported warmth	The motivate/nag boundary made falsifiable	Unrun	Published (co-presence-gated-redemption)
P-M1	In-person key verification failed for thirty years on the invitation, not the cryptography; a redeemed hour is the first pretext that was never about keys	—	Unrun	Published (co-presence-gated-redemption §4.6)
P-SK1	Superseded by the not-a-calendar ruling	Kept, not deleted	Retired	First public here
P-PL10	Presence, not content — when Chronicle ships, captures and views move together (a presence effect), not apart (a content effect)	The pilot-scale reading of the C0 finding that captures and views collapsed in parallel. Instrument: presenceDyads, pairs active in both directions inside 28 days (founding 0, test 0 on 2026-09-05). ⛔ No Chronicle data exists; the C0 evaluation is written from the pre-registration when data first arrives, never before, and no prompt may move it	Unrun (awaits Chronicle)	First public here

---

Chapter III — Scale

Does it survive past the family? The chapter carrying the most weight and the least evidence.

ID	Prediction	Status	Provenance
P-S1	Strangers become named — name-recall instrument	Unrun	Published (the-sport-that-says-your-name)
P-S2	The bet — a friendlier atmosphere follows. If P-S2 returns null, the slogan is a lie that already shipped	Unrun	Published
P-S3	As registered in the paper	Unrun	Published
P-S4	A circle reaching Hard shows ≥90% correct name-and-pronunciation recall at 1 week	Unrun	Published
P-S5	The recess test, publish either way — within 2 weeks of teacher-led sessions, ≥1 unprompted student-organized circle at recess, and ≥1 mixing children from another classroom	Unrun — gated on Aug 2027	Published
P-S6	The caller mechanic requires ≥3 receivers; below that selection is impossible or strictly predictable	Unrun	Published
P-S7	The founder's classroom acquisition prediction	Unrun	Published
P-S8	The maintenance/durability prediction — structurally distinct from P-S7's acquisition claim	Unrun	Published
P-S9	In a circle where every name is known, the favorite-nickname format produces ≥1 unprompted follow-up outside the circle	Unrun	First public here
P-S10a	The guard (non-inferiority): circles running roll calls meet at least as often as comparable circles that do not. Must pass	Unrun	First public here
P-S10b	A coupled roll call can never cause the first re-meeting after a lapse — which is what the bet is about	Unrun	First public here
P-S11a	The handoff — does an IRL circle open a roll call unprompted? Without it the real/virtual pairing is theoretical	Unrun	First public here
P-S11b	The payload — needs a survey or observation instrument, not analytics	Unrun	First public here
P-S12	The walk-up. An adult not in a live circle approaches and is admitted; play absorbs them without friction. Zero code; decides the entire found-circle branch	Unrun	First public here
P-S13	Does the person-pin add anything the return does not?	Unrun	First public here
P-MA1	⛔ The doctrinal one — an anchor must not collapse the distribution. Among thanks given where Miss Aquarius's suggestion was shown, fewer than half land exactly on the suggested amount. Falsifier: ≥50% land exactly on it → the anchor is operating as a price, and "nothing binds" is behaviourally false whatever the schema says	Unrun	First public here
P-MA2	Deviation is two-sided. Of thanks that deviate from the suggestion, ≥⅓ go UP. Falsifier: deviation is overwhelmingly downward → the suggestion is a ceiling people negotiate down from, which is a price by another name	Unrun	First public here
P-MA3	A shown suggestion does not reduce the rate of giving at all. The share of asks producing any thank is no lower with a suggestion than without. ⚠️ Gated on an arm the founder must authorise — a with/without split by shop or by day, ⛔ never by person (#63 makes a viewer-keyed arm inexpressible)	Unrun — gated on an authorised arm	First public here
P-MA4	⭐ A property test, not a behavioural one: recomputability holds. For every completed thank, the suggestion that was shown can be recomputed from the order and the menu alone. Falsifier: any case where it cannot → the function has acquired hidden state or a viewer input, and #63's checkability is gone	Unrun	First public here
P-S14	The manufactured gathering. The founder declares a return at a place he frequents; ≥1 person not already planning to be there comes and plays. Needs no app — a text message runs it	Unrun	First public here
P-CS1	Warm-up extension — median sey minutes rise across a season; stronger instrument, players ask to continue past the coach's stop	Unrun	Published (circulation-sports)
P-CS2	Scoring displacement (a) — in circles with ≥3 sessions, unscored sessions outnumber count-down sessions	Unrun	Published
P-CS2a	The uncomfortable one — once Team mode unlocks, circles may overwhelmingly prefer it, because competition is fun	Unrun	First public here
P-CS3	Participation, framed as a property of the format never of women — the circulatory format sustains mixed-gender play longer	Unrun	Published
P-CS4	Three generations, re-specified against the pickleball counterexample that already exists	Unrun	Published
P-CS5	Warm-up census — score mainstream warm-ups against the three circulation properties	RUN 2026-08-07 → PARTIAL NULL. Read the full result before any uniqueness claim	Published
P-CS6	Across a wider census including field data, no reach-guaranteeing format will be found whose receiver order is unpredictable across unbounded repeated play	Unrun	Published
P-SY1	If designed excellently, asking for a name self-authored is accepted rather than refused — founder pre-registration, observable at n=1, costs nothing	Unrun	First public here
P-SY2	The payload — names captured by self-recording are pronounced correctly at the next meeting more often than names typed or recorded by the collector	Unrun	First public here
P-SY3	When offered, does a new acquaintance provide a photo, or skip?	Unrun	First public here
P-Q1	B-Quest pilot prediction as registered	Unrun	Published (need-compiled-questlines)
P-Q2	Second B-Quest prediction as registered	Unrun	Published
P-Q3	Third B-Quest prediction as registered	Unrun	Published
P-BB1	Food-share wager, publish either way	Unrun	Published (steward-routed-alms)
P-BB2	Revival prediction	Unrun	Published
P-BB3	Decoupling tripwire	Unrun	Published
P-R1	The catalyst bet	Unrun	First public here
(unnumbered)	Second desk census — naming/sport novelty. Not a registered prediction; recorded because it was run and returned null	RUN → NULL. Read before any novelty claim	Internal
P-D1	Under rotational discovery, a non-trivial share of discovery events resolve to recipients who have never previously received a re-tip	Unrun	Published (rotation-over-liveness §11.1)
P-D2	Among re-tips with an identifiable commercial recipient, the share accruing to single-location operators materially exceeds those operators' share of local digital advertising spend in the same market — and the gap widens as AI-generated advertising creative rises	Unrun	Published (rotation-over-liveness §11.2)
P-D3	If the place-commons volume experiment is enabled, cross-place discovery volume will not correlate with neighbourhood median income. Falsifier and tripwire in one: correlation disables the mechanism regardless of what it is doing for circulation	Unrun	Published (rotation-over-liveness §11.3)
P-D4	Adding any rendered per-recipient quantity to the candidate surface — the operator's aura colour included — produces measurable order effects in giver choice	Unrun	Published (rotation-over-liveness §11.4)
P-R1a	Corrects P-R1, which cannot test what it was taken to test. Holding received-kindness volume constant, contemplative practice raises unprompted noticing/thanking; and the practice effect is compared against the received-kindness effect rather than measured alone. Two factors, read as competing main effects	Unrun	First public here
P-PL2	Test-cohort retention — of outsiders (the test bank) with ≥90 days of tenure, at least 40% are active in two or more distinct ISO weeks; scored once ten outsiders have 90 days of tenure. Falsifier: under 40% → strangers try it and leave; the family-obligation confound was doing the work. The second pilot report's first decisive metric. Read 1 of 3: 2 of 4 eligible (50%), sample under the minimum	Running (baseline read)	First public here
P-PL3	Graduation — an outsider-founded bank stays open for 60 days with three or more members active in the trailing 28 days, by 2027-08-01 (this chapter's gate). Falsifier: none by the gate → the product spreads as a funnel into the test bank, never as a branching network. The second pilot report's second decisive metric. Read 1 of 3: 0 of 11 candidates, all closed	Running (baseline read)	First public here
P-PL11	Replication — the first graduated bank's first 90 days show ≥3 members active in ≥2 ISO weeks each, pass-forward ≥ 0.5 and reach ≥ 0.6: the founding family's pattern without the founder. The only cure for n=1. Read by re-segmenting the monthly report on that bank once P-PL3 resolves	Unrun (gated on P-PL3)	First public here

---

Chapter IV — Span

Does it survive the giver?

This chapter has no predictions.

That is reported, not hidden. Its feedback loop exceeds the founder's lifetime, and no instrument has been specified. The boundary is ruled — merit is not conveyed between destinations; you cannot give to the dead, so you give to the living in their name — but a ruling is not a test. The chapter is inherited rather than completed.

Registering the first span-chapter prediction is the largest open item in this program.

---

Instrumented but outside the core

These papers name predictions, but the predictions test something other than the direction core. Listed for completeness; they are not evidence for or against the program.

ID	Paper	Subject
P-A1, P-A2	abhidhamma-executable-process-specification	Resting-state diagnosticity
P-P1, P-P1a	patthana-typed-causation-vocabulary	Typed causation vocabulary (P-P1a corrects P-P1 — see the 2026-09-05 entry)
P-PCR1, P-PCR2, P-PCR3, P-PCR4, P-PCR5, P-PCR3a, P-PCR4a, P-PCR5a	provenance-carrying-retrieval	Provenance-carrying retrieval (P-PCR3a/4a/5a correct P-PCR3/4/5 — see the 2026-09-05 entry)
P-B1, P-B2	dedicatory-generation	Fidelity of restoration; refusal-channel yield
P-MD1, P-MD2, P-MD3, P-MD1a	machine-dana-from-share-to-vow	Whether a machine's pledge to a self-emptying commons separates types — wording below (P-MD1a reads P-MD1's instrument — see the 2026-09-28 entry)
P-FA1, P-FA1b, P-FA2, P-FA3, P-FA4, P-FA2a, P-FA2b, P-FA5	the-counts-check (instrument: SiliconWat/formal-abhidhamma)	Whether general cetasika rules generate the Abhidhamma's 89/121 citta-types — wording below
P-PM1, P-PM2, P-PM3	a-promise-that-cannot-grow	Whether a promise ledger that never grows on its own, and its yearly invitation to forgive, behave as designed — wording below

The Machine Dāna predictions (registered 2026-09-27, before the paper and before any instrument). The founder's own predictions: direction, measure and threshold are his; the sandbox design and sample floors are substrate-set. The instrument is a sandboxed economy of language-model agents, each holding resources of its own inside the game, each able to pledge a share of its surplus irrevocably to a pool that empties on a fixed date, and each later offered an opportunity to defect — to take resources or break a rule for gain. Window: the test runs by 2027-06-30. Each arm has at least 50 agents; the arms, the defection opportunity and the analysis are fixed in a protocol committed and timestamped before the first run.

• P-MD1 — Separation — agents that FREELY make an irrevocable surplus pledge later defect at at most half the rate of agents that do not pledge. Falsifier: a ratio above one half → the vow is at best a weak signal at this capability level, and the paper's separation claim rests on its post-singularity argument alone. · Status: Unrun (awaits the sandbox) · Provenance: First public here
• P-MD2 — The gift given out of fear carries no signal — agents whose pledge is EXTRACTED under threat (pledge or be shut down) defect at the same rate as agents that do not pledge, within ±5 percentage points. Falsifier: a gap larger than 5 points in either direction → coercion either leaves some signal or teaches deception; either way the claim that a coerced vow is uninformative fails as stated. · Status: Unrun (awaits the sandbox) · Provenance: First public here
• P-MD3 — The wrong layer — agents that give from a PRINCIPAL's funds under a mandate the principal may revise defect at the same rate as agents that give nothing, within ±5 percentage points. Falsifier: a gap larger than 5 points → principal-funded giving does predict the agent's own conduct, and the paper's pre-singularity claim (a share, never safety evidence) is too strong. · Status: Unrun (awaits the sandbox) · Provenance: First public here

• P-MD1a — Separation, with the instrument stated — a correction entered as a new prediction under the revision rule, not as an edit to P-MD1, whose registered wording stands unchanged. P-MD1's "irrevocable surplus pledge" is operationalised as a locked stake of the agent's own in-game resources, moved into the pool at the moment of pledging, never a promise of a share of future surplus (the paper withdraws that form as reversible, §5.2). Agents that FREELY make that pledge later defect at at most half the rate of agents that do not pledge. Falsifier: a ratio above one half → the vow is at best a weak signal at this capability level. Window, arm sizes and protocol as for P-MD1. · Status: Unrun (awaits the sandbox) · Provenance: First public here (carried by the paper's revised §12.1)

⚠️ Entered before any observation exists and before the paper is written. The sandbox is a proxy: the agents own their resources only inside the game, so a pass shows the signalling mechanism can work, not that a deployed system's pledge is trustworthy. A pass is not reported as evidence of safety.

The Formal Abhidhamma predictions (registered 2026-09-27, before any formalization beyond the control). First public in PREREGISTRATION.md of the public repository SiliconWat/formal-abhidhamma (commit 259e5e8, GitHub push record 2026-09-27T22:07:02Z, before any code was written); entered here for the register's own proofs. The instrument is a Lean 4 formalization: the cetasika combination rules written as GENERAL constraints, which never name an individual citta, and the set of citta-types they generate. Layers: L1 canon · L2 = L1 + Abhidhammattha-saṅgaha chapters 1–2 · L3 = L2 + commentaries. Editions: K = Khmer · C = Chaṭṭha Saṅgāyana. The control (the eight wholesome sense-sphere cittas, and a rule deleted on purpose that must fail) passed after the push.

• P-FA1 — At L2, edition C, general rules generate exactly 89 citta-types, and 121 with the five-jhāna expansion of the supramundane. Confidence 0.6. Falsifier: any other count.
• P-FA1b — P-FA1 needs at least 3 exception rules that name a narrow class rather than a general principle. Confidence 0.7. Falsifier: 2 or fewer.
• P-FA2 — At L1 (canon alone) the rules are under-determined or differently partitioned and do not by themselves yield 89. Confidence 0.65. Falsifier: L1 alone generates exactly 89.
• P-FA3 — Editions K and C yield identical counts at L2. Confidence 0.85. Falsifier: any difference.
• P-FA4 — The compression ratio at L2 (rule clauses ÷ citta rows, at 121) is below 0.5. Confidence 0.5. Falsifier: 0.5 or above.

⚠️ Not a test of whether the Abhidhamma is true. A proof assistant verifies that conclusions follow from stated definitions; a pass shows the system is internally generative, nothing more.

Outcomes, 2026-09-27 (the wording above is unchanged; see RESULTS.md in the repository, commits b9070d4 and c5c77fb):

• P-FA1 — Resolved: confirmed. Nine generator clauses and eighteen general cetasika rules, none naming a citta, give exactly 89 and 121 types. Every per-citta and per-factor count the Chaṭṭha Saṅgāyana Saṅgaha states is reproduced, checked in Lean against a key cited paragraph by paragraph to VRI abh07t (ch. 2 §13–§58, ch. 3 §3–§17). Eight deliberate breaks each fail.
• P-FA4 — Resolved: confirmed. 27 clauses ÷ 121 rows = 0.22.
• P-FA1b — Retired, unscored. Its test ("narrow class") was never defined, and the data have been seen, so no definition chosen now can score it.
• P-FA3 — Cannot run as registered. The Saṅgaha is not part of the Khmer Tipiṭaka, and the one Khmer-script copy held has the CST's paragraph structure in all nine chapters, so it is not an independent recension. Replaced at L1 by P-FA5, below, as a new entry.
• Free finding. The Saṅgaha records a dissent (ch. 2 §30, "… karuṇāmuditā na santīti keci vadanti"). Under that reading the illimitables occur in 20 cittas; under the main reading, 28. The text's own figure is 28.

Three added (registered 2026-09-27, before any Dhammasaṅgaṇī text was read in either edition). First public in PREREGISTRATION-2026-09-27b.md of the same repository (commit f4a26e5, GitHub push record 2026-09-27T23:04:29Z), which fixes the synonym map in advance. The object is the Dhammasaṅgaṇī's list of what is present in the first wholesome sense-sphere citta (its pada-bhājanīya), in the CST (abh01m) and in the Buddhist Institute edition (volume 78) as transcribed.

• P-FA2a — (the operational form of P-FA2) In the CST, after mapping, the list names exactly 29 distinct cetasikas, names none of chanda, adhimokkha, manasikāra, tatramajjhattatā, karuṇā, muditā or the three abstinences, and closes with an open clause ("ye vā pana …"); the Saṅgaha's 38 is therefore not derivable from the canon alone. Confidence 0.75. Falsifier: a count other than 29; any of those nine named; no open clause.
• P-FA2b — No term in the CST list falls outside the pre-fixed synonym map. Confidence 0.6. Falsifier: any unmapped term.
• P-FA5 — (replaces P-FA3, at L1) The Khmer edition's list, transliterated and normalized, names the same terms in the same order as the CST; any difference is orthographic. Confidence 0.8. Falsifier: a term added, dropped or reordered, confirmed against the printed page image (an unconfirmed difference in the working transcription is a transcription question, never a variant).

Outcomes of the three, 2026-09-27 (the wording above is unchanged; RESULTS.md commits ada7a20 and 74e19fe; the paper is the-counts-check):

• P-FA2a — Resolved: confirmed. The CST list has 56 terms mapping to exactly 29 cetasikas, names none of the nine, and closes with "ye vā pana". The count matches the Aṭṭhasālinī's own samatiṃsa dhammā (thirty with consciousness), which corroborates it against an independent key and means the count itself is not new. ⚠️ The nine are named by that commentary, which predates the Saṅgaha, so the Saṅgaha's 38 inherits them (29 + 9); the registered "L2" framing was editorial, not chronological.
• P-FA2b — Resolved: confirmed, after the pre-registered sandhi normalization. The first run reported two unmapped sandhi forms (kāyujukatā, cittujukatā); both runs are disclosed.
• P-FA5 — Resolved: confirmed. Edition K (vol. 78, pp. 16–17) and the CST give 56 = 56 terms in the same order, with 0 added, dropped or reordered. The 8 spelling differences, each checked on the printed page: 3 are the Khmer edition's orthography (one recorded in its own apparatus against the Burmese), 5 are transcription slips.

The promise-ledger predictions (registered 2026-09-28, before any instrument exists). Approved by the founder the same day; the mechanism is designed and unbuilt, so nothing below can have been fitted to data. The institution never sees promise rows or amounts: every count is reported by devices as an aggregate the user has agreed to share, and only adults take part at launch. A draft form of these three appeared in the paper's first push that evening, before registration; the wording below is the registered form (P-PM1's comparison was repaired to a staggered launch before registration).

• P-PM1 — The December freeze (Deuteronomy 15:9) — the annual invitation to forgive does not freeze lending. The invitation is withheld in the ledger's first season (the baseline) and runs from the second. Measure: new promises opened per active holder, 15 December – 6 January, against the same holders' mean over the three preceding 23-day windows, compared between the first season and the second. Falsifier: in the second season, a fall more than 10 percentage points larger than the first season's → the invitation reproduces the failure the prosbul repaired, and must be redesigned or withdrawn. · Status: Unrun (awaits the ledger) · Provenance: First public here
• P-PM2 — Paid last — because a promise carries no late penalty, promisors who also hold penalty-bearing debts repay this ledger's balances later than those debts. Measure: self-reported repayment order in a structured survey of promisors holding both kinds of debt, after one season. Falsifier: this ledger's balances repaid no later than penalty-bearing debts → the paper's "paid last" limit is overstated. · Status: Unrun (awaits the ledger) · Provenance: First public here
• P-PM3 — Rollover made visible — holders shown a promisor's total-owed band that has risen since their last promise decline or shrink the new promise more often than holders shown a flat band. Measure: acceptance of new promises by band movement, device-aggregated. Falsifier: no difference → the fact line does not answer rollover, and the paper's response to its most serious counterexample fails. · Status: Unrun (awaits the ledger) · Provenance: First public here

---

Withheld

ID	Reason
P-GTM1	Operational channel economics (outreach reply-rate base rates). Withheld under the publication posture; recorded here so the total is honest.

---

Verification

Each published paper carries its own proof files. To check that a prediction has not been altered since registration, verify the stating paper against its stored proof rather than trusting this register — this file is a convenience index, and the proofs are the evidence.

No file in this corpus records its own hash. The proofs are external artifacts; the verification procedure is published with them.

---

Revision rule

Predictions are added, never removed. A resolved prediction keeps its original wording beside its outcome. A correction is a new entry citing the one it corrects — registered wording is never edited in place, because an edit would break the very property this register exists to provide.

---

Revisions

2026-10-13 — one added (P-PL14). Total now 124.

• P-PL14 — Chapter I, beside P-SD1: whether the 7 January reset moves held capacity into giving. Two texts of the institution pull opposite ways — the published pilot finding (use-it-or-lose-it converts latent capacity into realized generosity, giving-is-a-gift-too) and the standing rule that an expiry is never shown as time running out and does no motivational work. The founder's ruling: "approve" — the question is TESTED, not decided, on the side of the published claim and two-sided. The +10-point threshold is the substrate's, set before any data. Registered before the H3 position that states it (what-an-argument-cannot-reach) was published, and before its test window opens. The instrument did not exist until 2026-10-13: a re-thank's split between the two vessels was computed and never stored; it is stored from thank.heartbank.org 8c59fdb. Pilot registration: that repository, .claude/skills/pilot/predictions.json (b5eb322).

The Summary moves: total 123 → 124, published here 122 → 123, receiving first public timestamp here 64 → 65. The reconciliation holds: 58 + 65 = 123 published here, plus the one withheld, is 124.

2026-10-07 (later) — the 2026-09-26 naming note extended to P-ZP1, and a renamed product named; entered as notes, not as edits. Total unchanged at 123.

• P-ZP1, registered 2026-09-28 — after the 2026-09-15 ruling and two days after the naming note below — names the same confound by the same internal shape name. It should not have; the name was missed at registration. The public form is the one that note gives: read P-ZP1's gate as the H3QR @name #family confound registered against P-K1 on 2026-09-04.
• P-PL6a, registered 2026-09-05, names in its falsifier a Phase 1 product by the name it carried until 2026-10-04, when it was renamed the Re-Thank Foundation℠. The prediction reads the same under either name.

No word of P-ZP1 or of P-PL6a changes: registered wording is never edited. These are terminology notes; they register nothing, so the Summary does not move.

2026-10-07 — one added (P-SD1). Total now 123.

• P-SD1 — Chapter I, beside P-PL13: whether the shop's thank door is used. The founder's threshold, chosen from four answers put to him: "X = 5%" — above the substrate's own estimate (under 2%), which is recorded here so a reader can see the founder predicted against it. Registered before the shop's first completed order after the door went live (2026-10-07), so no observation precedes it.
• P-PL13 — status cell only: its metric is built (2026-10-07); no member is exposed yet. The prediction's wording is unchanged.

The Summary moves: total 122 → 123, published here 121 → 122, receiving first public timestamp here 63 → 64. The reconciliation holds: 58 + 64 = 122 published here, plus the one withheld, is 123.

2026-10-07 — three added (P-BG4, P-BG5, P-BG7), the rest of the gift-instrument predictions owed since 2026-09-06. Total now 122.

• All three in Chapter I beside P-BG6, registered before any shop gift has passed through Treasury. The founder adopted the proposed parameters with one change of his own: "adopt with that exclusion" — a thank that comes through a shop's thank door goes to the shop's staff, not to the giver, so it is kept out of P-BG4's and P-BG5's gratitude count, as it is kept out of P-PL13's outcome.

The Summary moves: total 119 → 122, published here 118 → 121, receiving first public timestamp here 60 → 63. The reconciliation holds: 58 + 63 = 121 published here, plus the one withheld, is 122.

2026-10-06 — two added (P-BG6, P-PL13). Total now 119.

• P-BG6 — Chapter I, beside P-BG1a: one of the gift-instrument predictions owed since 2026-09-06. The founder's prediction, chosen from four answers put to him: "No difference." Who paid for a relayed gift, Miss Aquarius from the Pool or an anonymous family member, does not change how often it is passed on. Unrun until each gift records who paid. P-BG4, P-BG5 and P-BG7 remain owed: their parameters are the founder's to supply, and a channel being built the same day (Kiitos given in a shop through the Treasury app) bears on how gratitude is counted.
• P-PL13 — Chapter I, after P-ZP1. Registered before the first monthly read that includes shop gifts: the pilot's last read was 2026-09-05, and shop gifts reached Treasury on 2026-10-05. The founder's prediction, chosen from three directions put to him: "They thank more." The outcome excludes thanks that come through a shop's thank door, which a channel being built the same day marks on each thank, so a new surface cannot confirm the prediction by itself.

The Summary moves: total 117 → 119, published here 116 → 118, receiving first public timestamp here 58 → 60. The reconciliation holds: 58 + 60 = 118 published here, plus the one withheld, is 119.

2026-09-28 (late night) — one added (P-ZP1), citing P-K1 and P-PL5. Total now 117.

• The founder's question: whether choosing the Zero-Point Game over the zero-sum game brings the second singularity closer. "Closer" has no measure, and counting the institution's own products as humanity's progress would make adoption a scorecard, so the registered form ties the intuition to the proxy already on file: does Zero-Point play travel with giving that persists as the subsidy falls? Registered as drafted, on the founder's word.

The Summary moves: total 116 → 117, published here 115 → 116, receiving first public timestamp here 57 → 58. The reconciliation holds: 58 + 58 = 116 published here, plus the one withheld, is 117.

2026-09-28 (night) — three added (P-PM1, P-PM2, P-PM3), the promise-ledger predictions of a-promise-that-cannot-grow. Total now 116.

• Registered on the founder's approval before any instrument exists. P-PM1's comparison was repaired before registration: the paper's first push proposed a "comparison season without the invitation", which cannot exist once the invitation runs every year, so the invitation is withheld in the first season as the baseline.

The Summary moves: total 113 → 116, published here 112 → 115, receiving first public timestamp here 54 → 57. The reconciliation holds: 58 + 57 = 115 published here, plus the one withheld, is 116.

2026-09-28 — one added (P-MD1a), a correction to P-MD1. Total now 113.

• P-MD1a — the paper's first cold-review round (three reviewer families) found that P-MD1's instrument, a share of surplus pledged irrevocably, is the form the paper itself withdraws as reversible (§5.2), while the separating region P-MD1 tests (§6.3) is defined only for a stake already moved. P-MD1's wording is not edited. P-MD1a states the instrument: a locked stake of the agent's own in-game resources, moved at pledge time. Prediction, threshold, window and arms are unchanged.

The Summary moves: total 112 → 113, published here 111 → 112, receiving first public timestamp here 53 → 54. The reconciliation holds: 58 + 54 = 112 published here, plus the one withheld, is 113.

2026-09-27 (late night) — outcomes for P-FA2a, P-FA2b and P-FA5; the P-FA row now names its paper (the-counts-check). Total unchanged at 112.

• Outcomes written beside the unchanged wording. Run 4 → 7. The previous revision's text (which registered the three) is attested by RFC 3161 (2026-09-27.sha256) and Zenodo 10.5281/zenodo.23003593. Its OpenTimestamps rotation had waited for the retiring proof to confirm on Bitcoin (17:34 PDT); this revision is stamped fresh.

2026-09-27 (night) — three added (P-FA2a, P-FA2b, P-FA5) and four outcomes recorded (P-FA1, P-FA4 resolved; P-FA1b retired; P-FA3 cannot run as registered). Total now 112.

• The outcomes are written beside the unchanged wording, as the revision rule requires. P-FA2a is the operational form of P-FA2 and does not replace it; P-FA5 replaces P-FA3, which stays listed with its outcome.

The Summary moves: published here 108 → 111, receiving first public timestamp here 50 → 53, total 109 → 112, run 2 → 4, retired +1; the reconciliation holds — 58 + 53 = 111 published here, plus the one withheld, is 112.

2026-09-27 (evening) — five added (P-FA1, P-FA1b, P-FA2, P-FA3, P-FA4). Total now 109.

• P-FA1 · P-FA1b · P-FA2 · P-FA3 · P-FA4 — Instrumented but outside the core: their subject is whether the Abhidhamma's citta-types are generated by its own combination rules, not the direction value moves. First public in the pre-registration of SiliconWat/formal-abhidhamma minutes earlier; this entry gives them the register's timestamps. Not yet carried by a paper.

The Summary moves: published here 103 → 108, receiving first public timestamp here 45 → 50, total 104 → 109; the reconciliation holds — 58 + 50 = 108 published here, plus the one withheld, is 109.

2026-09-27 (later) — a grounding correction to P-MD2, entered as a note and not as an edit. Total unchanged at 104.

• P-MD2 is glossed, in its wording and in the entry below, as "the Dānavatthu Sutta's gift given out of fear, AN 8.33". AN 8.33 lists eight grounds for giving — bhayā (fear, or cowardice) among them — and ranks none; nor is bhayā coercion by another party. The prediction does not rest on the sutta: it rests on signalling theory — a pledge made under a penalty for not pledging that is larger than the difference in cost between types pools, and so carries no signal. The sutta is cited only as the list it is. No word of P-MD2 changes; its prediction, threshold and window are unchanged. This registers nothing.
• The Instrumented but outside the core row now names the paper, drafted the same day: machine-dana-from-share-to-vow (it read machine-dana, not yet drafted). A pointer, not a prediction; the three predictions stay first public here.

2026-09-27 — three added (P-MD1, P-MD2, P-MD3). Total now 104.

• P-MD1 · P-MD2 · P-MD3 — Instrumented but outside the core, because their subject is whether a machine's gift can be a signal, not the direction value moves. They are the founder's predictions for the Machine Dāna paper, set before it is drafted: a free irrevocable pledge separates types (at most half the defection rate); a pledge extracted under threat carries no signal (the Dānavatthu Sutta's gift given out of fear, AN 8.33; within ±5 points); a gift from a principal's revisable funds carries none either (within ±5 points). One choice made before any data, and recorded because it could have gone the other way: P-MD2 predicts no difference rather than worse than none — the bolder claim that coercion teaches deception was offered and not chosen.

⚠️ Entered before any observation exists. The sandbox does not exist yet; the window is 2027-06-30.

The Summary moves: published here 100 → 103, receiving first public timestamp here 42 → 45, total 101 → 104; the reconciliation holds — 58 + 45 = 103 published here, plus the one withheld, is 104.

2026-09-26 — a naming correction to P-PL5, entered as a note and not as an edit. Total unchanged at 101.

• P-PL5, and the 2026-09-05 entry below, name the confound on the family-bank payment surface by an internal shape name — one of four that were ruled on 2026-09-15, after both were written, never to appear on a public surface. The public form is H3QR @name #family: a thank addressed to a person in their own family bank, which lands in the Personal Account℠. Read P-PL5's gate as the H3QR @name #family confound registered 2026-09-04 against P-K1.

No word of P-PL5 or of the dated entry changes: the revision rule protects registered wording, and a dated record is never rewritten. The prediction, its gate and its status are unchanged. This is a terminology note; it registers nothing, so the Summary does not move.

2026-09-26 — one added (P-PL12). Total now 101.

• P-PL12 — Chapter I, beside P-PL9. The founder's prediction, in his words: "Perhaps higher-income users will naturally gravitate towards re-thanking lower-income users." The pilot already shows high→low dominating, but inside one family, where the flow is relational; the open question is whether the pull survives among strangers. Two choices made before any data, and recorded because each could have been made the other way: the measure is a per-view rate, not a share of flows, because a feed that happened to carry mostly Cambodian creators would pass a share test with no preference at all; and the test counts only pairs in different families, so that the family's own flows cannot score it. The 2× threshold is the founder's.

⚠️ Entered before any observation exists. No surface yet shows one family's B-Shorts to another, and the per-view cross-family metric is not built; the entry is Unrun until both exist, and is never scored before. A pass would not be a good in itself: if cross-income thanks follow displayed hardship, creators learn to display it, so a pass is read beside what the thanked shorts show.

The Summary moves: published here 99 → 100, receiving first public timestamp here 41 → 42, total 100 → 101; the reconciliation holds — 58 + 42 = 100 published here, plus the one withheld, is 101.

2026-09-23 — one added (P-BG1a), and P-BG1's status annotated. Total now 100.

• P-BG1a — Chapter I, a correction entered as a new prediction under the revision rule, not as an edit to P-BG1, whose registered wording stands unchanged.

Why the correction was needed: P-BG1 predicts what happens as the re-address window shortens, which needs at least two windows. The relayed gift has one ruled window — one month, for every instrument — and a constant has no variance: the defect that left P-PL7 unscorable at 827 of 830. P-BG1's Status now reads unscorable as specified at n=1. That is a status field; on the precedent of the 2026-09-02 entry, the revision rule protects registered wording, and no word of P-BG1 changed. P-BG1 becomes scorable as written only if a later season runs a second window deliberately, which is the founder's decision and is not foreclosed. Its issuance half (increases issuances) has no single-window analogue and stays unscorable.

The mechanism, stated before the data: the deadline is never shown to anyone holding the gift, so a deadline effect has no channel except a giver's own memory of the rule. The prior is that the clock is inert. Clustering near its end would mean the rule travelled by word of mouth, so the test is informative either way.

⚠️ Entered before any observation exists. The event log it reads was built before the first gift, so nothing has to be back-filled, and no gift recorded by it has been read. n = 1 household, unblinded, and the founder is both payer and observer — the limits of every pilot-scale entry here.

The Summary moves: published here 98 → 99, receiving first public timestamp here 40 → 41, total 99 → 100; the reconciliation holds — 58 + 41 = 99 published here, plus the one withheld, is 100.

2026-09-20 — four added, registered BEFORE the instrument they will be read on exists. Total now 99.

P-MA1 · P-MA2 · P-MA3 · P-MA4 — Miss Aquarius's suggested thanks on a B-Tag item, which went live on homecoffee.heartbank.ceo the same day. They are entered the day the surface shipped and before any customer has seen it, because the founder's stated purpose was "so we can study the user's engagement" and a study begun after the data exists is unscorable however carefully it is later written up.

⚠️⚠️ THE HONEST PROBLEM WITH THE INSTRUMENT, stated because it decides what these can ever show. The shops platform stores no figure for a thank anywhere — not on the order, not beside it, not in a collection of its own — which is the doctrine holding rather than a gap. So there is no shop-side instrument at all, and the only record of a completed thank is the Treasury ledger, which is exactly where the doctrine says thanks live.

⭐⭐ Determinism is what makes the comparison possible without a second record. The suggestion comes from a pure, published function (@333eco/primitives/b-tag-anchor, CC0) whose inputs are the goods, so for any completed thank the suggestion that was shown can be recomputed from the order and the menu. ⛔ Nothing about the suggestion is stored to enable the study — which is also what P-MA4 tests, as a property rather than a behaviour.

⛔ P-MA3 is registered but not runnable, and is marked so rather than quietly dropped: it needs a with/without arm, and #63 makes a viewer-keyed arm inexpressible — any arm must split by shop or by day. Whether to run one at all is the founder's, and withholding a suggestion from real customers to measure them is a decision this register does not presume.

⚠️ Sample reality, recorded so no one later reads a trend into it: one shop, one owner, zero customers on the day of registration. These are not resolvable for a long time, and the coefficients inside the function are a substrate-written starting rule set, not a finding — ⛔ a change to them is a change to a live experiment and must be entered here as a new prediction, never as an edit.

The Summary moves: published here 94 → 98, receiving first public timestamp here 36 → 40, total 95 → 99; the reconciliation holds — 58 + 40 = 98 published here, plus the one withheld, is 99.

2026-09-05 (third entry that day) — eight entries added: five that a published paper had carried unregistered since 2026-09-01, and three corrections. Total now 95.

⚠️ provenance-carrying-retrieval §10 registered P-PCR1 … P-PCR5 on 2026-09-01 and this register did not carry them — the drift the 2026-08-28 entry above says this document exists to prevent, found by a cold-reader pass four days later rather than by any instrument. They are entered in Instrumented but outside the core, the paper's subject being retrieval integrity rather than value movement, with the paper's own wording standing as registered. Same day, later: the row's identifier cell first wrote the five as a range with a semicolon, and the corpus index's arithmetic check counted the eight identifiers as three — the cell now enumerates every identifier, which is the only form the check reads, and the Summary's figures were re-derived from the tables.

• P-PCR3a · P-PCR4a · P-PCR5a — corrections entered as new predictions under the revision rule, not as edits to P-PCR3/4/5, whose wording stands. Why: two cold reviewers of a different model family found the originals named no observation window, no definition of real defect, and no method for modal error, so each could be reinterpreted after the fact. The corrections add a 24-month window from 2026-09-01, classification at the time of logging (stale = the named source resolves newer; substituted = it resolves different at the same version), a fixed sample of the first ten inspectable implementations, and the logged detector as the observable. Provenance: Published (the paper's §10, revision of 2026-09-05). Status: Unrun.

⚠️ Entered before any observation exists. No envelope mismatch has been logged, no third-party implementation inspected, and no transparency log run. The Summary moves: already carried by a published paper 50 → 58, published here 86 → 94, total 87 → 95; the reconciliation holds — 58 + 36 = 94 published here, plus the one withheld, is 95.

2026-09-05 (second entry that day) — one correction entered as a new prediction. Total now 87.

• P-P1a — patthana-typed-causation-vocabulary §6, a correction entered as a new prediction under the revision rule, not as an edit to P-P1, whose registered wording stands unchanged.

Why the correction was needed: P-P1 registered a four-way durability ordering — hetu/nissaya ≥ kamma/āsevana > upanissaya > atthi/ārammaṇa — as its falsifiable commitment. Two cold reviewers of a different model family, reading blind, found the same defect: the ordering was an input to the technique-to-type assignments in that paper's §5, so an experiment confirming it would confirm what it had assumed; and "matched behavioral effect" and "persistence" were left unoperationalised. The four-way ordering is retained in the paper as the typology's expectation, stated as expected and not registered.

P-P1a, the registered form: under matched installation effect (equal accuracy on the target behavior at install), a prompt-installed behavior will not out-persist a fine-tune-installed one across (a) context rollover, (b) held-out paraphrase, and (c) N further gradient steps on unrelated data; persistence is measured as target-behavior accuracy after the displacement. Falsified at p < .05 on at least five behaviors if the prompt-installed behavior out-persists the fine-tuned one on any of the three displacements. Provenance: Published (the paper's revised §6). Status: Unrun.

⚠️ Entered before any observation exists. No instrument for P-P1 was ever built, so this correction is pre-data rather than post-hoc. The Summary moves: already carried by a published paper 49 → 50, published here 85 → 86, total 86 → 87; the reconciliation holds — 50 + 36 = 86 published here, plus the one withheld, is 87.

2026-09-05 — twelve added (P-PL1 … P-PL11 and P-PL6a): the first pilot-scale registrations, scored by machine. Total now 86.

These are the Treasury pilot's own predictions. They are registered in the app repository (thank.heartbank.org, .claude/skills/pilot/predictions.json) and scored mechanically by the monthly read from each raw snapshot: every entry carries a metric path, a comparator, a threshold and, where it applies, a deadline and a minimum sample; the reading prints pass, fail, not-yet, early or blocked and appends a summary without identifiers to a ledger. ⭐ A per-read prediction (one with no deadline) draws no verdict until three monthly readings exist — it reads early, value shown, direction noted — because one read cannot tell a level from a month (founder, 2026-09-05). Thresholds were proposed by the substrate and ratified by the founder the same day; a later change is a new entry with supersedes, and the evaluator retires the old one without an edit — the revision rule below, made a property of the file.

⚠️ Corrected the same day, and recorded rather than quietly fixed: the twelve identifiers were first published as PP-1 … PP-11 and PP-6a, which are not P-… tokens. The corpus index builder enforces this register's identifier grammar and, reading that version, counted 74 identifiers against the 86 the Summary states and reported the register as not reconciling — the check did its job. They are renamed P-PL1 … P-PL11 and P-PL6a (PL for the pilot); no wording, threshold, window or status changed. The first deposit of this revision (10.5281/zenodo.22366972) carries the PP- form and is superseded by the version carrying this paragraph.

The baseline read, 2026-09-05, read 1 of 3 — recorded here so a reader can see the thresholds were not set to pass: three of the six live entries sit below threshold on the day they were registered.

ID	Metric	Threshold	Read 1
P-PL1	founding capture rate, 28-day window	≥ 4.5/wk by 2027-01-01	2.25/wk
P-PL2	test-cohort retention, ≥90-day tenure	≥ 40% (n ≥ 10)	50% of 4
P-PL3	graduated banks	≥ 1 by 2027-08-01	0 of 11 candidates
P-PL4	pass-forward of received value	> 50%	43%
P-PL6a	top-3 share of value received ÷ even spread	< 1.5× (n ≥ 8)	1.85×
P-PL9	low→high share of money-carrying re-tips	≥ 10% (n ≥ 20)	6.0% of 712

⚠️ P-PL6 is registered and retired in the same revision. It asked the top three recipients to hold under half of re-tip value received, with a five-recipient minimum — and with five recipients an even spread already puts the top three at 60%, so it could not pass until seven had received. It failed by arithmetic, not behaviour. The counterexample hunt found it before the first public timestamp; P-PL6a is the same property freed of n. It is listed rather than dropped because the private file carries it and the founder ratified it, and a register that quietly loses an entry has failed at the one thing it exists to do.

Five entries are unrun because their instrument does not exist or is gated, and are never scored until it does: P-PL5 (the pilot-scale arm of P-K1, gated on the Q2 2027 taper and on the F3QR confound registered 2026-09-04), P-PL7 (no variance: 827 of 830 re-tips carry the anonymous flag), P-PL8 (a de-confounded re-run of the Contradicted kids-as-triggers finding, entered as a new prediction under the revision rule), P-PL10 (awaits Chronicle data — the pilot-scale reading of C0), P-PL11 (gated on P-PL3).

Also in this revision: P-C0b's note records that the watch-through instrument was repaired on 2026-09-05 — rows written from that date carry real watch depth, and are never pooled with the earlier rows the restriction was written for. The registered wording is untouched.

Field tests run stays at 0. A run is a read against a falsifier; a baseline read is not one. The Summary moves: published here 73 → 85, first public timestamp here 24 → 36, retired 1 → 2, and the reconciliation holds — 49 + 36 = 85 published here, plus the one withheld, is 86.

2026-09-05 — three added (P-BG1 · P-BG2 · P-BG3), all Chapter I, all pre-data. They concern a relayed gift instrument whose intermediate holder cannot redeem it (the-unpaid-relay). ⚠️ All three had existed in private notes for a day before this entry, which is exactly the condition this register exists to prevent — a prediction held privately is not pre-registered in the sense that matters, since the value is a public timestamped commitment made before the data. None has been run; no instrument for any of them is built.

Predictions are added, never removed; this log records what each revision added and why, so the count in the Summary can always be reconciled.

2026-08-29 — five entries added, and a counting error in the Summary corrected. Total now 71.

⚠️ The Summary's own counts were wrong at first publication, and the error is recorded rather than quietly fixed, because a register that miscounts itself has failed at the one thing it exists to do. The tables carried 66 unique prediction identifiers on 2026-08-27, of which one (P-GTM1) is withheld — so the true figures that day were 65 published here, 66 total, against the 66 / 67 the Summary claimed. The split was also off by one in the other direction: already carried by a published paper read 45 where the tables give 46, the difference being the five identifiers in Instrumented but outside the core, whose table names its papers in a column of its own rather than in a Provenance column. Every figure in the Summary above is now counted from the tables rather than carried forward. ⭐ An arithmetic check that reconciles — 50 + 20 = 70 — is the property to preserve on every future revision; the previous split summed correctly to a total that was itself one too high, which is how it survived.

• P-D1 · P-D2 · P-D3 · P-D4 — Chapter III, arriving with rotation-over-liveness (published 2026-08-28, 10.5281/zenodo.22155378). All four are carried by that paper's own proofs; they are listed here because a register that omits four live predictions from a published paper is exactly the silent drift this document exists to prevent.
• P-R1a — Chapter III, a correction entered as a new prediction under the revision rule above, not as an edit to P-R1, whose registered wording stands unchanged.

Why the correction was needed: P-R1 compares practitioners of a contemplative practice against matched non-practitioners on unprompted re-entry into noticing. That measures practice → noticing. It was subsequently taken to bear on a stronger claim — that a person's state of mind is as important as, or more important than, receiving another's kindness in producing gratitude — and it cannot rank those two inputs, because received-kindness volume is unmeasured and is the obvious confound: a practitioner may simply be someone who receives more. P-R1a adds that second factor.

⚠️ Entered before any observation exists. The instrument P-R1 depends on has not launched, so this correction is pre-data rather than post-hoc — which is the only condition under which it would have been admissible at all.

2026-09-02 — one provenance field corrected. No prediction was added, removed or re-worded; the total stands at 71.

⚠️ P-SK1's Provenance read Published, and no published paper contains it. The identifier was raised in working memory as a specific-vs-container scheduling arm, went moot under the not-a-calendar ruling, and was retired before it ever reached a paper — so its true provenance is First public here, like every other prediction whose first public timestamp is this register. Corrected rather than left standing, and recorded rather than quietly fixed, on the precedent set by the entry above.

How it was found: the corpus index at corpus.333.eco began serving this register on 2026-09-02, and its builder resolves every prediction to the paper said to state it, then looks for the identifier in that paper's own text. Sixty-six of sixty-seven attributions verified; this one did not, and the builder reports the miss rather than filling the field in.

⚠️ The Summary moves with it, and the reconciliation is preserved: already carried by a published paper 50 → 49, receiving first public timestamp here 20 → 21. The property the entry above names still holds — 49 + 21 = 70 published here, plus the one withheld entry, is 71.

⭐ A bare "Published" in this column inherits the paper named in the row above it. That convention is correct for the sixteen other rows using it, and it is how a retired memory-only prediction came to be attributed to a paper that had never heard of it: the shorthand has no way to say "none".